b70-optimization-lab

Laguna routed gather/finalize Phase-A preflight terminal

Date: 2026-07-24 America/Toronto

Status: terminal before runner; no retry, component evidence, endpoint, or submission.

Outcome

The exact M=8 post-W2 gather/finalize candidate completed Stage 0 and received one immutable Phase-A timing/exactness authorization:

• tooling commit: 1bc3db422daefd2c5e7fe915eaff8dfd850ec920;
• packet-only commit: 180826bea272c73e6cf767df1b02fc0b80ef018a;
• authorization packet SHA-256: 383ca2f804793047be3898a3e12a4772d6e5592385b0483b1edd6d0af529a39a;
• candidate _moe_C SHA-256: 6a6794249421aceb51f14980a3e2c0b0a9d7b492abf2f8d25b129b86f099bc5b; and
• fixture SHA-256: 0b1ea43d0a724cc64eaf6636b99076afd852846f79c06b4db264f2a511689259.

The first and only authorized execution completed its five registered discovery probes, then stopped on the first strict-idle sample at 2026-07-24 05:04 America/Toronto. The campaign root was never acquired. The durable failure is:

/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/
gather-finalize-phase-a-1bc3db422-v1-preflight-failure.json

Its SHA-256 is bd91f212e4728cd410ea68df043394789ef903dcdf7245907553e6054dbc6d4b and its terminal status is component_failed_stop_before_runner.

Root cause

The frozen coordinator required the JSON result of xpu-smi ps -j to be exactly:

{"process_list":[]}

The installed tool instead emits a top-level device_util_by_proc_list. A read-only diagnostic immediately after the failure contained four rows for the querying xpu-smi process itself, one per card, and no model or service process. That process had exited by the subsequent host-process check. The diagnostic raw SHA-256 was f47866b297acef6711ab08e9ee7d15d783d2d9b9492e537731b2ccaf2528ec7d.

This was a harness-schema failure, not evidence that the candidate was inexact, slow, or even imported. The fail-closed result is nevertheless binding. The packet authorized one campaign and explicitly set retry_authorized=false; it must not be rerun, replaced, or rescued after observing the failure.

What did and did not run

Ran:

• five read-only xpu-smi discovery -j mapping probes; and
• one read-only xpu-smi ps -j strict-idle sample.

Did not run:

• candidate or incumbent native-module import;
• Torch/XPU tensor allocation;
• any candidate, control, downstream, timing, or profiler primitive;
• model load, endpoint, generation, cold prompt, or benchmark;
• counter capture, payload construction, network access, LocalMaxxing submission, service change, or reboot.

The candidate therefore remains unmeasured. No performance or correctness conclusion may be drawn from this terminal preflight.

Prevention and next action

Future one-shot packets must test their idle-parser grammar against a retained capture from the installed tool before authorization. A future coordinator should launch xpu-smi with Popen, retain the child PID, accept only observer rows bound to that exact PID and process name, and fail on every other process row. Mock-only schema tests are insufficient.

The next Laguna experiment must be materially distinct and separately preregistered. The current approved record remains 33.89498511171744 tok/s, LocalMaxxing cmrx6p5dv001bo4017hb7sixz, at vLLM 8936aac144929190c1e53f8b8624ca397ce16f5b plus XPU kernels b6076ce1249ffee0e30bee528f4cd15c3bffb234.

This site is open source. Improve this page.